---------- Forwarded message ----------
From: 王昊然 <msl0000023...@gmail.com>
Date: Thu, 20 Oct 2022 00:33:18 +0800
Subject: Re: Bug#1022049: libnettle8: Illegal instruction on IBM POWER7
To: Niels Möller <ni...@lysator.liu.se>
2022-10-20 0:18 GMT+08:00, 王昊然 <msl0000023...@gmail.com>:
>> The ppc specification says "vmrgow is treated as a Vector
>> instruction in terms of resource availability.", it's not entirely clear
>> to me what that means, and if checking for altivec support should be
>> enough.
>>
>> The fat setup for ppc is intended to enable the crashing code path only
>> if the feature bits PPC_FEATURE_HAS_ALTIVEC and PPC_FEATURE_HAS_VSX are
>> both set in the status word returned from
>> ...
>
> Thanks for the hint, I guess this issue is likely caused by my kernel which
> wasn't have CONFIG_VSX enabled. But in my understanding, this feature is an
> extension to AltiVec, not the AltiVec itself, so nettle shouldn't assume
> VSX
> to be available if they seen AltiVec available.
>
>
>> You can get some diagnostics from the initialization process by setting
>> the NETTLE_FAT_VERBOSE environment variable, and override the automatic
>> detection with the NETTLE_FAT_OVERRIDE environment variable.
>
> $ NETTLE_FAT_VERBOSE=1 test-program
> libnettle: cpu features:
> libnettle: enabling altivec code.
> Illegal instruction
>
>
>> Can you check what getauxval(AT_HWCAP) returns on your system?
>
> whr@debian:~/src$ cat cpu-feature-test.c
> #include <sys/auxv.h>
> #include <stdio.h>
>
> int main() {
> unsigned long int hwcap = getauxval(AT_HWCAP);
> printf("hwcap = 0x%lx\n", hwcap);
> return 0;
> }
> whr@debian:~/src$ gcc -Wall cpu-feature-test.c
> whr@debian:~/src$ ./a.out
> hwcap = 0xdc0065c2
>
>
>> Is the program that crashes running under a vm, or is the kernel running
>> on the bare metal? Each layer of vm tends to be ian opportunity to
>> introduce
>> errors in the list of available processor features.
>
> Operating systems on this machine are always running on the Power
> Hypervisor,
> which is a part of the server firmware.
>
I just checked the nettle source code, it is indeed correctly checked both
PPC_FEATURE_HAS_ALTIVEC and PPC_FEATURE_HAS_VSX; and with my HWCAP has
PPC_FEATURE_HAS_VSX set, I think it hits a limitation of this processor
feature checking logic: hardware supporting it, but the kernel didn't.
I will try to build a new kernel with CONFIG_VSX enabled by tomorrow.
